[Question] WaveRNN Vocoder about multilingual_text_to_speech

[Tomiinek]

Hello, I am sorry for a late response.

1. I had problems with training the model too and it was very very slow. You can always try out different vocoders like MelGAN or WaveGlow (I do not know what is SOTA currently).

2. I have two questions. Do the training audios have the same scale and are they normalized to the same volume? Are the quant values up to 1000 ok?

[sooftware]

I'll check the scale of audio. Is the frame length of your tacotron Mel spectrogram 50 ms and frame shift 12.5 ms?
Also, Is there a WaveRNN that I learned with Lj speech?

Thank you for your response.

[sooftware]

Audio's scale is the same. (up to 1000)

[sooftware]

Furthermore, if I insert the pre-built mel (from gta.py) into the WaveRNN you opened as input, it produces no speech at all.

[Tomiinek]

Is the frame length of your tacotron Mel spectrogram 50 ms and frame shift 12.5 ms?

Yes, it is.

Also, Is there a WaveRNN that I learned with Lj speech?

I think that it is in the original repository that I forked.

Audio's scale is the same. (up to 1000)

Ok, and is the scale the same if you apply some volume normalization such as sox infile.wav outfile.wav gain −n −3? I would really expect audio scale in [-1, 1].

Furthermore, if I insert the pre-built mel (from gta.py) into the WaveRNN you opened as input, it produces no speech at all.

I do not understand, can you please explain in more detail?

[sooftware]

Hmm... Thank you for your answer!!
Never mind the last question. I'll think about it a little more.